Dopp is a proxy for Jev-compatible decision APIs. Point your app's base URL at it instead of Jev: the requests and replies stay the same, and every request is kept. Label them, or pick an LLM or any other service to label them for you. Then train a small open model on them on your own CPU and let it answer, with Jev as the fallback. The model also runs offline or in a browser tab.

Quickstart

See the whole loop in about three minutes, with no keys or accounts. You need Node 22+ and uv:

git clone https://github.com/doppsh/dopp && cd dopp
npm install
npm run example:home-assistant

It starts a local Dopp with a throwaway database and a stand-in service, sends Home Assistant voice commands through the proxy, trains Tiny on your CPU, downloads the offline folder and asks it every command. Add -- --keep to look around the dashboard afterwards. The first run also installs PyTorch and downloads bge-small (about 130 MB).

To use it with your own app, run it on your machine and point your app at it. If you'd rather not run anything, dopp.sh is the same product, hosted.

How it works

Jev is TypeSafe's hosted decision API (POST https://api.typesafe.ai/v1/systemone). You send a state (text or any JSON) and a few typed questions about it: choice (pick one option), noul (yes or no) and score (a point on a scale). You get back an answer to each, with a probability for every option. Any server that takes and returns that shape is "Jev-compatible".

What Dopp adds:

• Every request is kept as a training example, in a request log that shows who answered, how long it took, and why a fallback fired.
• Routes and setups. Each key belongs to a route. The route's setup decides who answers (Jev, another Jev-compatible service, an LLM, or a model you trained), what happens when one is unsure or slow, and who labels the requests a model learns from: you by default (fix or confirm an answer on the Requests page), or an LLM or any upstream you pick.
• Train small open models on your own requests: Tiny (bge-small, 34 MB int8) on your own CPU in a few minutes, or Laya, GLiNER and Kev on GPUs. Each version is measured against what answers today on requests it never saw.
• Run them where you want: as an offline folder that answers /v1/systemone with no network (a laptop, a Raspberry Pi), in the browser (WebGPU), or hosted on your own GPUs (Modal).

The numbers in the animation are from the repo's own Home Assistant example (npm run example:home-assistant, measured on a laptop, 30 Sep 2026): 222 voice commands in the HA-Jev shape with the label written next to each one, a Tiny model trained on 205 of them on the CPU in about 60 s. On 68 held-out answers it agrees 100% of the time, at 1.8 ms a request. Those are template commands, so it's an easy test; the same loop on real commands from a real house is written up, with its numbers, at dopp.sh/blog/run-jev-offline.

This repo is the whole self-hosted product: the proxy and API (a Cloudflare Worker you can run locally), the dashboard, the trainers, the browser runtime and the offline server. It's built for one operator. There's no sign-up: you set a password and make the keys your apps use.

It's early. The same loop runs end to end on every commit, in CI, as the Home Assistant example.

Run it on your machine

You need Node 22 or newer. You don't need a Cloudflare account or a GPU.

npm install
cp worker/.dev.vars.example worker/.dev.vars

In worker/.dev.vars, set:

• ADMIN_PASSWORD: the dashboard's password, at least 12 characters.
• KEY_SECRET: any long random string (openssl rand -hex 32). It encrypts keys you paste into the dashboard.
• Who answers: TYPESAFE_API_KEY (your TypeSafe key), or JEV_URL pointing at any Jev-compatible server. Without one of these, requests are kept but nobody answers them.

Then:

npm run dev

This builds the dashboard, creates a local database (SQLite, in .wrangler/) and serves everything at http://localhost:8787.

1. Open http://localhost:8787 and sign in with ADMIN_PASSWORD.

2. Getting started makes your first route and shows its key (us_…), with a request you can paste into a terminal.

3. Send a request:

   DOPP_KEY=us_...
   curl http://localhost:8787/v1/systemone \
     -H "Authorization: Bearer $DOPP_KEY" -H "content-type: application/json" \
     -d '{"state":"My card was charged twice.","questions":{"refund":{"type":"noul","instructions":"Does the customer want money back?"}}}'
   

4. Open Requests: the request is there, with who answered and how long it took.

npm test checks all of this end to end, against a stand-in for Jev and a throwaway database, with no keys.

Point your app at it

Change the base URL from https://api.typesafe.ai to your Dopp server, and the API key to a route key. To have TypeSafe bill your own key for a request instead of the server's, also send it as the x-jev-key header (forwarded, never stored). The API is documented at /docs.md on your server and in /openapi.json.

Only keys made on your server get in. Anyone holding one can spend the upstream keys set on your server, so treat them like API secrets and revoke any you've shared from the route's page.

Train on your own machine

Tiny trains on your CPU, with no GPU and no Modal account. It takes a few minutes for a few hundred requests on a laptop. With uv installed:

1. Add to worker/.dev.vars:

   TINY_ENGINE_URL=http://127.0.0.1:8788
   SELF_URL=http://localhost:8787
   MODAL_SECRET=<any long random string>   # the Worker and the trainer check it on every call
   

2. In a second terminal: npm run trainer. The first run installs PyTorch and downloads bge-small (about 130 MB).

3. Restart npm run dev. On Models, Tiny is ready, priced at $0, on this machine's CPU.

A trained version downloads from the Models page as an offline folder; see Offline folder. Without uv, use Python 3.10 to 3.12: pip install -r engines/requirements-tiny.txt, then python engines/tiny_local.py.

To see the whole loop without touching the dashboard, run an example:

• Home Assistant: voice commands in, an offline folder that answers them. npm run example:home-assistant
• The Gatekeeper: a castle guard's brain, trained here and run in a browser tab with browser/tiny.js. npm run example:gatekeeper

Train with your own Modal

Laya, GLiNER and Kev train on GPUs, on Modal, on your account. Without it Dopp still proxies, logs, labels, compares and trains Tiny on your CPU, and the dashboard says which engines aren't deployed on your server.

1. pip install modal && modal setup

2. Make the shared secret the engines and the Worker use to trust each other (pick any long random value):

   modal secret create dopp-engine UNDERSTUDY_SECRET=<value>
   

   and set the same value as MODAL_SECRET (in worker/.dev.vars, or npx wrangler secret put MODAL_SECRET -c worker/wrangler.toml).

3. Deploy the engines you want, from the repo root:

   Engine	Command	Trains on	Serves on
   Laya (and Laya multilingual)	modal deploy engines/modal_engine.py	H100	T4
   Tiny (bge-small + one head per question)	modal deploy engines/modal_tiny.py	T4	your machine (offline folder)
   GLiNER (small, Decide)	modal deploy engines/modal_gliner.py	A10G	T4
   Kev (0.8B, 4B)	clone github.com/jaredpalmer/kev to kev-repo/ (or set KEV_ROOT), then modal deploy engines/modal_kev.py	H100	T4
   Open models, untrained (JevK5)	modal deploy engines/modal_open.py	none	L4

4. Put the URLs Modal prints into worker/wrangler.toml (ENGINE_URL, TINY_ENGINE_URL, GLINER_ENGINE_URL, …; the commented lines show which is which) and restart.

5. Set SELF_URL in worker/wrangler.toml to an address Modal can reach. After training, Modal sends the version's files (the offline folder, the browser copy) back there. A server on localhost can still train, but those files can't come back until it has a public address: deploy it (below), or run a tunnel such as cloudflared tunnel --url http://localhost:8787 and use the address it prints.

Modal bills you by the second for GPU time; the Usage page shows what each run and each awake hour cost at Modal's list prices. Engines scale to zero when idle.

Offline folder

A trained Tiny version downloads as a folder (model, tokenizer, tiny.json, serve.py) from the Models page. It answers POST /v1/systemone like Jev, with no network, on any machine with Python (Linux ARM included). Inside the folder:

pip install onnxruntime tokenizers numpy
python serve.py --port 8799

It listens on port 8787 unless told otherwise, the same port as npm run dev, hence --port above. serve.py is a copy of engines/tiny_serve.py. To have Dopp ask it, set JEV_URL=http://127.0.0.1:8799/v1/systemone.

In the browser

A Tiny offline folder also runs in a browser tab: browser/tiny.js (with browser/wordpiece.js, no other dependencies) loads it with ONNX Runtime Web and answers like serve.py (same tokens and choices; probabilities within a few hundredths). The Gatekeeper example shows it in a bare page.

Trained Laya versions also get a browser copy (ONNX, run on WebGPU by dashboard/public/laya-web.js), which the Worker serves only to you. Like the offline folder, it is sent back to SELF_URL after training.

Deploy to your Cloudflare account

npx wrangler login
npx wrangler d1 create dopp                 # put the id in worker/wrangler.toml
npx wrangler r2 bucket create dopp-files    # keep it private
npm run deploy:db
npx wrangler secret put ADMIN_PASSWORD -c worker/wrangler.toml
npx wrangler secret put KEY_SECRET -c worker/wrangler.toml     # and TYPESAFE_API_KEY and the others you use
npm run deploy

Set SELF_URL in worker/wrangler.toml to the address it's served at. The dashboard asks for ADMIN_PASSWORD, so pick a long one: wrong guesses are slowed down, but only within each Worker instance. To put the whole dashboard behind your own login as well, Cloudflare Access works in front of it; leave /v1/systemone open to your apps.

Configuration

Secrets (names only; see worker/.dev.vars.example): ADMIN_PASSWORD, KEY_SECRET, TYPESAFE_API_KEY, MODAL_SECRET, GEN_API_KEY (with GEN_BASE_URL and GEN_MODEL for a writer other than Gemini), OPENROUTER_API_KEY.

Vars (worker/wrangler.toml): SELF_URL, JEV_URL, PRICE_MARGIN (cost shown = list price times this; default 1), the engine URLs, REQUESTS_PER_MINUTE_PER_KEY (default 120), and RESPONSE_META=off to leave Dopp's understudy block out of replies for clients that reject unknown fields.

Layout

Models and their licenses

Dopp downloads base weights from Hugging Face when an engine is built; this repo contains no weights.

Python packages the engines install: laya (Apache-2.0), gliner2 (Apache-2.0), PyTorch, Transformers, ONNX Runtime. Jev is TypeSafe's hosted API and is not part of this repo.

What dopp.sh adds

The hosted version adds accounts and sign-up, GPUs and provider keys already set up, one bill instead of your own provider accounts, and hosting subscriptions for trained models. Those parts aren't in this repo.

License

MIT, see LICENSE.

---

Made by @justkrup. Hosted version: dopp.sh.

_{The name is short for doppelgänger: a double of a living person. Here, a model of your own that answers your app's questions, on your hardware.}